Domain Atlas / Caseworker documentation & copilots
CDTFA Axyom Assist
California's tax agency ran the first state generative AI call-center assistant through a complete governed procurement arc: a competitive sandbox in which two vendors were each paid one dollar to test for six months in a secure environment, a 10-month pilot in a simulated environment, and production rollout to roughly 375 agents completed August 22, 2025 under a 12-month, 445,000 dollar contract. The pilot projected a minimum 1.5 percent per-call time saving, and by May 2026 the agency had declined to renew the contract, with the official who launched the project saying the working system 'didn't quite save as much time as we had hoped.'[4]
What happened
California's call-center GenAI project began as one of the first proofs of concept under Governor Newsom's Executive Order N-12-23 (September 6, 2023), the statewide framework mandating a deliberate evaluation-then-deployment process for generative AI in state government. The California Department of Tax and Fee Administration announced its solicitation on January 16, 2024, through the RFI2 "Request for Innovative Ideas" vehicle, framing the goal as helping agents search reference materials to answer taxpayer phone and live-chat inquiries; about 100 potential vendors attended the informational session. The scale is what made the case attractive: the department fields more than 800,000 taxpayer inquiries a year across phone, chat, and email (over 660,000 calls, 17,000 emails, and 127,000 chat messages in FY2022-23 alone), with roughly 375 call-center agents inside a 3,696-employee department. Peak tax season brings up to 10,000 calls a day — about four times normal — with waits stretching from roughly 4 minutes to 20, historically absorbed by reassigning about 280 temporary staff from other units. In May 2024 the state awarded sandbox contracts to two competing vendors, Ignyte Group and SymSoft Solutions, each paid $1 to test for six months in a secure, isolated sandbox overseen by the California Department of Technology, using publicly available data only, with the Office of Data and Innovation designated to run a human-centered final assessment. The launch-time safeguards were explicit: mandatory human review before any AI-generated response reached a caller, a contractual vendor obligation to monitor and report on response accuracy, coherence, and appropriateness, and employee training on detecting inaccurate output. Consumer Watchdog's Justin Kloczko publicly warned at launch that GenAI "makes mistakes" and that call-center staff may lack the qualifications to verify AI accuracy.
After a 10-month pilot in a simulated environment, CDTFA down-selected to SymSoft Solutions' Axyom Assist — a retrieval-augmented assistant built on a commercially hosted large language model — and moved it into production from March 2025 under a 12-month, $445,000 contract, an arc the Governor's office announced on April 29, 2025 as a first-in-the-nation deployment. The production tool was internal-facing only: it transcribed live calls in real time, detected the taxpayer's need, retrieved candidate answers from more than 16,000 CDTFA web and policy pages spanning 42 tax and fee programs, let agents ask ad-hoc questions, and generated post-call summaries, with agents reviewing everything before use (vendor-described capabilities, agency-confirmed scope). Departmentwide implementation completed on August 22, 2025. The benefit figures attached to the project are pilot projections, not production measurements: call-center chief Thor Dunn told trade press that the simulated environment "did show some potential to save at least one and a half percent time on some of those calls," which the agency extrapolated to roughly 100,000 minutes per year — capacity for about 10,000 additional calls annually. No published production-environment measurement, methodology, or ODI assessment report has been located. Department leadership framed the deployment as a scaling proof point for other state call centers.
By May 2026, CDTFA had declined to renew the contract. Government Operations Secretary Nick Maduros — the department's director when the project began — put the reasoning plainly: "Even though the solutions worked in practice, they didn't quite save as much time as we had hoped." The department shifted to a similar off-the-shelf tool from Amazon Web Services already covered by an existing, larger contract, reasoning that commodity tools had caught up with the custom build during its own contract term. The non-renewal sat inside a broader governed portfolio review: of eight executive-order-driven GenAI pilots totaling more than $5.8 million, three others also ended — a housing tracker that never left pilot, an inspection tool that lost fiscal-year funding, an agency that chose internal development over the vendor product — while contracts at three other departments continued. A separate CDTFA proof of concept, the DATRPS project — a roughly $900,000 Deloitte Consulting machine-learning contract for tax-return processing — was reported by Tax Notes in December 2024, from documents obtained under the California Public Records Act, to have been hindered by human error; the error was a state technology team member's miscalculation in the vendor-evaluation scoring system during procurement, and the episode belongs to that sibling project, not to the call-center assistant, its data, or its corpus.
The sociotechnical reading
Most oversight columns in this atlas are drawn hopefully: an evaluation body that receives reports and issues findings somewhere off to the side of the loop that matters. This case is the counterexample the domain needed, because here the oversight chain terminates in a purchaser-side reviewer that demonstrably acted — twice. The competitive sandbox eliminated one of two vendors before production; the renewal review ended a working deployment because a projected benefit did not justify a known price. Neither firing was scandal-driven. There is no plaintiff, no exposé, no adjudicated failure anywhere in this record — the governance lever that did the work was the most ordinary one government owns, the procurement calendar, wired so that the deployment had to re-earn its contract with numbers on a schedule. The map makes the mechanism visible: vendor accuracy reporting was contractual (the vendor grading its own product, a real but conflicted check), the evaluation team received it, and the renewal authority above them held a genuinely credible exit — a one-year term, a commodity substitute already under contract, and a manual practice that never stopped working. Credible exit is the load-bearing element. The department could decline renewal without a capability cliff precisely because the copilot had been layered onto a working process rather than replacing one, and because commoditization pressure — the force that outran the custom build mid-contract — cut in the purchaser's favor.
The honest complications travel with the praise. First, the benefit was never production-verified: the 1.5% per-call saving and its 100,000-minute extrapolation are projections from a simulated pilot, self-reported by the agency through trade press and echoed in vendor marketing, and the human-centered assessment charged to the state's data-innovation office at award has no published report — so the arc's celebrated return-on-investment judgment was, on the public record, a decision made about projections, by the same institutions that produced them. The Lab diagram carries that as a low-rung cross-check pathway — the vendor's contractual monitoring of its own answers, and a human-centered assessment charged at award whose report has not been located — with nothing independent of the deployment behind either. Second, the design's most important feature is an absence: the knowledge corpus is publicly available data only and is never fed by the assistant's own outputs, so the classic copilot contamination loop — model output entering the store that grounds the next model output — is deliberately unwired, leaving AI-drafted call summaries in internal workflow records as the one residual write channel. An absence like that is a decision someone must keep re-making through every future integration, which is why the map draws it as a latent edge rather than leaving it off the page. Third, the launch-time automation-bias warning was never resolved, only bounded: mandatory review and training stand in front of every answer, but the concern about whether surge-season staff reassigned from other units can meaningfully verify answers across 42 programs concentrates exactly where peak load makes checking most expensive. The distinct lesson for the Field Guide is that the strongest governance artifact in this domain is not a dashboard or an audit — it is a contract short enough to lapse, attached to an evaluation that must be repeated, in a market with a substitute. What the record does not establish matters too, and the boundary holds as always: nothing here measures what callers experienced, and the tool's effect on the taxpayers at the other end of 800,000 annual inquiries was never the quantity any part of this system computed.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library.